Why we’re building
Grey Materia.

Dear customers,

Our team brings experience from Meta, Google, and AWS, along with executive leadership at Lambda and in public cloud. We helped build AI cloud as it grew, and learned a lot about what customers need from it: dependable capacity, fast access to data, and systems that work together.

We’re applying that experience to physical AI. We see an underserved need for infrastructure that supports robotics simulation, data generation, model training, evaluation, and deployment. These workloads have requirements beyond training and serving an LLM.

Physical AI infrastructure requirements.

Physical AI shares several requirements with LLM development. Large robotics models and world models need GPU memory, fast networks, and high-throughput storage. Physical AI also requires simulation, sensor processing, and repeated interaction with an environment. The balance of those workloads changes as a model develops.

That changes the hardware mix. Photorealistic sensor simulation needs rendering-capable GPUs. Policy training can combine parallel physics simulation with learning, without needing photorealistic images for every task. Large model training has a different compute profile. A useful cloud has to match each job to the right resources and coordinate the work between them.

Robot learning uses camera recordings, sensor readings, and actions aligned in time. Teams need to know which data, environment, and model version produced a result. Synthetic data also needs checks for useful coverage and physical consistency. Storage, data preparation, and versioning have to support that work as datasets grow.

Evaluation runs throughout development. A policy needs to be tested across changes in objects, lighting, contact, and motion. Experience from real robots can feed further policy training, so results need to flow back into the next experiment. Keeping simulation, training, and evaluation connected is an infrastructure requirement in its own right.

Then there is the computer on the robot. A model has to work within its memory, power, and response-time limits. Optimizing or quantizing it changes what needs to be tested. Running a container successfully in the cloud is only one step; the model also needs validation on its intended hardware.

What we’re building for.

Cloud providers already support serious physical AI work. We see room to make the complete workflow easier to operate, especially for teams that would otherwise have to assemble and maintain the connections themselves.

We’re designing Grey Materia around that need: complementary DGX and RTX resources, a shared data layer, compatible software environments, and a path from simulation to validation on physical hardware. Your data and checkpoints should remain usable as the work moves between those stages.

We believe many enterprises will develop physical AI models for their own operations. Some will adapt existing models to their equipment and tasks. Others will train specialized policies, robotics foundation models, or world models. Their operating knowledge and data will shape what those models can do.

We want those teams to spend less time maintaining infrastructure and more time improving model performance.